You are correct. All the audio is prerecorded. Each recording is connected with the text it matches. Theoretically we would have separate audio files for qatlh and Qatlh. However, the system does not recognize them as separate text tags and will happily play the "qatlh" file for Qatlh and the "Qatlh" file for qatlh. The system does not recognize that they are different and in actuality will only save one such file. Each time we rerecord one, it overwrites the other, so at this time, we can only have one recording or the other. We are working on a solution and hope to do better soon.
I thought about this while I was recording today and then thought, "hey, what does Duolingo do about read/read (past/present) or record/record (noun/verb) or any other of dozens of spelled alike/pronounced differently words in English?" Why would Klingon be treated this way even if it thinks they are spelled the same? Russian, too, has numerous words that are spelled the same but stressed differently depending on meaning. Can we not tap into that?